for · 2018. 8. 17. · by Vivek Gopalan PennsylvaniaState University j...
Transcript of for · 2018. 8. 17. · by Vivek Gopalan PennsylvaniaState University j...
-
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
A Robust Optimization Approach for Selecting Urban Fire Engine Company Locations
by Vivek Gopalan
Pennsylvania State University | email: [email protected]
Sponsor: Dr. Murali Kodialam1
AbstractWhen a fire breaks out in a city, an emergency call to 911 is made and a fire engine
company responds to the incident. This average response time is an important system met-ric that must be kept under a suitable threshold. The goal of this project is to study the re-lationship between the number and locations of fire engine companies and their response times to help cities optimize resources. Since the number and locations of fires a re not known apriori, any selected set of engine company locations must be robust across a wide spectrum of fire i ncidents. F urthermore, i f t oo much e mphasis i s p laced o n optimizing resources and system costs, some engine companies could bear a disproportionate fraction of the workload, leading to dissatisfaction or fatigue among firefighters. This project there-fore also considers the incremental cost of ensuring equitable engine company workloads. A methodology combining integer programming techniques, ensemble learning, and local search using genetic algorithms is proposed to determine robust locations for the fire en-gine companies. Tested with a data set for the city of Philadelphia, the results support the hypothesis that optimizing fire engine company locations can result in significant savings for resource-strapped cities.
1. IntroductionOn July 5, 2014, four children were killed in a 3-alarm fire t hat b roke o ut in
Philadelphia [14] and a similar fire entailed evacuating a hundred residents from an apart-ment building more recently [17]. When a fire o r m edical e mergency o ccurs, a c all is placed to 911 and even a minute reduction in the response time (i.e., the time between the call and arrival of Emergency Medical Services (EMS) personnel at the site) could save several lives [18]. A recent audit report published by the Office of the Controller for the City of Philadelphia, reports that the National Fire Prevention Association (NFPA) has set a standard where the response time for the fire engines should be within 5 minutes and 20 seconds after the call is dispatched for 90 percent of their runs. The Philadelphia Fire Department responds to approximately 54,000 calls a year [18]. While response times can be reduced by adding more ambulance shelters or fire engine company locations, cities are already budget-constrained and must do their best with very limited resources.
The fields of operations research/management science have contributed immensely to improving urban public services [8]. While EMS encompasses many domains, such as locating ambulance shelters [1, 5], and optimizing police patrols [12], this research will focus on studying the relationship between number and locations of fire engine companies and their response times to emergency incidents.
1Address: Nokia Bell Labs, 791 Holmdel Road, Holmdel, NJ 07733.Email: [email protected]
Copyright © SIAM Unauthorized reproduction of this article is prohibited
244
-
Previous work has studied the relationship between travel distances and travel39times for fire engines, and has also outlined an algorithm for engine company reloca-40tion(i.e., when an engine company is called away to respond to a fire, another engine com-41pany may be relocated to its base as a surrogate, in case a fire breaks out), but has ignored42the development of robust solutions for fire engine company location problems [10, 11].43We focus on obtaining robust solutions for fire engine company location problems. A so-44lution to a facility location problem is considered robust if it performs well under a wide45range of changes in any stochastic parameters associated with the system. Baron et al. [2]46consider robust facility location when demand for the product produced by the facility47varies considerably over multiple time periods. Cui et al. [7] treat the case of facility dis-48ruption (e.g., a factory producing parts for a car unexpectedly goes out of commission) and49develop robust solutions for this problem. Expanding on previous investigations, this pa-50per finds optimal solutions to engine company location selection under spatial uncertainty51in demand (i.e. it is not known precisely where fires will break out and engine companies52must be located to accommodate a wide range of scenarios for fire locations).53
This paper addresses the following research questions:54
1. What is the minimum number of fire engine companies needed to provide satisfac-55tory coverage of fires in a given city?56
2. Given a required number of engine companies, where should these engine companies57be located?58
3. What resources should be made available at each fire engine company? In addition59to the standard pumper-type truck, fire companies utilize a more functional and ex-60pensive ladder truck. How many pumper and ladder trucks should be placed at each61location?62
To address these research questions, we propose a decision support system (DSS)63to determine engine company locations as outlined below and in Figure 1.64
1. This paper proposes a new metric for measuring robustness in the context of fire65engine company location. Robustness is defined in terms of a tuple (β, p), where at66least β percent of fires that break out in a fixed time period are covered with threshold67probability p (the probability of coverage p decreases with response distance (time)68traveled). More specifically, the research answers the following question: what is69the minimum number of fire engine companies needed within a city so that β % of fire70incidents can be covered with a threshold coverage probability p?71
2. Fire scenarios are generated via a Spatial Poisson process; the minimum number of72engine companies and their locations are identified via the integer program EC-PSCP73(Equity Constrained Probabilistic Set Covering Problem). This is Phase I in Figure 1.74
3. The engine company locations identified for each scenario are combined via an En-75semble Learning algorithm, using the notion of “voting” for the most effective facil-76ities (Phase II in Figure 1).77
245
-
4. The current paper also considers resource allocation at each location. A variety of78resource configurations may be deemed feasible at this stage, in the spirit of “solution79plurality” [9]. Resources (e.g., pumper vs. ladder trucks) are allocated to each location80by solving a Constraint Satisfaction Problem using Genetic Algorithms.81
Section 2 of this paper describes the data used and provides details of the solution82algorithm. Section 3 presents a discussion of computational results and Section 4 concludes83the paper with relevant public-policy recommendations and offers future research direc-84tions.85
246
-
2. Solution Approach86
2.1.A. Method Overview87
Figure 1 provides an overview of the procedure (algorithm) coded in Python,88which is illustrated using data obtained for the city of Philadelphia.89
90
91
Figure 1: Robust Optimization Process for Selecting Fire Engine Company Locations92
93
247
-
The algorithm (solution methodology) requires user-provided inputs and creates94certain outputs:95
1. User provided inputs: The user must provide a tuple (β, p), guiding the DSS to pro-96vide solutions where β % of fires can be covered with probability p. The coverage97probability itself depends upon the travel speed of fire engines and the distance be-98tween the fire and the responding engine company. An expert user must provide99a model for how the coverage probability varies with travel time (see Figure 2). In100the computations presented in this paper, a linear decay in the coverage probabil-101ity function is assumed. If a fire engine reaches a fire after t minutes, the coverage102probability is stated below (tmin and tmax are user-specified parameters, see Figure1032):104
p(t) =
1, t ≤ tmin
− t−tmintmax −tmin + 1, tmin < t ≤ tmax0, t > tmax
Response times less than tmin are deemed satisfactory and response times exceeding105tmax are unacceptable.106
107
108
109
110
111
112
113
Figure 2: Illustration of Two Coverage Probability Functions
Outputs created by the solution methodology: Number of engine companies needed, their arcGIS coordinates and the number and types of fire engines (pumper or ladder trucks) at each location.
248
-
This solution method, discussed below, consists of the five parts (refer to Figure1141).115
2.1.B. Data116
As a case study to illustrate the algorithm, we consider the city of Philadelphia.117The city is home to about 1.5 million residents and can be geographically partitioned into118155 neighborhoods (Figure 3). For each of these neighborhoods, a public-domain arcGIS119data set was obtained (see Figure 3) identifying the neighborhood population and the neigh-120borhood centers (i.e., the x and y coordinates for the centers in the arcGIS frame of refer-121ence). The precise spatial locations of fire incidents (for a 1-year period) were not available122for this project, but from public-domain information, it is known that a total of 54485 fire123incidents occurred during 2015 [18].124
2.2 Methods125
2.2.1 Generating Candidate Engine Company Locations126
The city of Philadelphia consists of 155 neighborhoods (as shown in Figure 3127below) and for each neighborhood, a square grid is drawn, with the center of the square128aligning with the center of the neighborhood, thereby masking the whole city (see top box129in Figure 1). Each corner point of the grid is assumed to be a potential engine company130location and the mesh size for the grid is a parameter that can be controlled by the user. The131finer the mesh size, the closer the location problem is to a “continuous location” problem.132
249
-
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
Figure 3: Map of Philadelphia Neighborhoods [16]
2.2.2 Generating Fire Location Scenarios
A scenario is simply a set of locations for fires that occur during a fixed period. If a data set were available with the exact spatial locations of fires for a year, the contin-uous spatial distribution to approximate the discrete set of real world fire events (the ob-served data) can be fit using Kernel Density Estimation (KDE), which is a non-parametric approach to density estimation from a set of discrete points [6, 13, 4]. The smoothing parameter (bandwidth) determines the extent of spread of each point. Once the spatial distribution is known, fire scenarios can be generated by sampling from this spatial distri-bution. However, fitting a spatial distribution for fire incidents is not the main objective of this paper. Moreover, fire incident location data were unavailable for this project, although an aggregate number for the total number of fires in the city of Philadelphia was available. The total number of fires was allocated to each of the 155 neighborhoods in proportion
250
-
to the neighborhood’s population. Fires were then randomly generated for each neighbor-148hood. One complete fire scenario consists of a set of spatial locations for fires in each of149the 155 neighborhoods. The scenario generation method repeatedly creates fire scenarios150for the city of Philadelphia. The next step is to solve for the best engine company locations151for each scenario using the method outlined in the next sub-section (2.3).152
2.3 Solving the Equity Constrained Probabilistic Set Covering Problem (EC-PSCP) for153each Scenario154
The Equity Constrained Probabilistic Set Covering Problem (EC-PSCP) repre-155sents the core of the methodology presented in this paper. An EC-PCSP is solved once156for each generated fire scenario (and the solutions later aggregated via an Ensemble Algo-157rithm (see section 2.4)). The EC-PCSP is an integer program that is a variant of the classical158set covering problem. As a preamble to solving the EC-PCSP, a set covering matrix A must159be developed. The matrix A has one row for each fire incident and one column for each160potential engine company location. A matrix entry aij is 1 if an engine company at location161j can cover a fire at i. The following steps enable the computation of matrix A.162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
Step 1: The Euclidean distance between each candidate facility j (node of the grid) and each fire incident location i in the scenario is computed.
Step 2: The Euclidean distance is converted to a travel distance using a DETOUR INDEX (equal to the ratio of actual travel to Euclidean distance) of 1.42. Prior research supports the use of such an index [9, 3].
Step 3: Using the user-specified travel speed for fire engines, the travel time matrix T is computed. T has (i, j)th entry = t(i, j) = (travel distance between i and j)/(engine speed).
Step 4: Using the travel time, compute the coverage probability p (as in Figure 3) and deter-mine if a facility at j can cover a fire at i . Develop a set covering matrix A, with (i, j)th entryaij = 1 if the coverage probability ≥ user-specified threshold probability and 0 otherwise [Note: The user-specified threshold probability is the second field in the tuple (β, p)].
Step 5: Use the set covering matrix to develop a mathematical programming formulation for the fire engine company location problem. The set covering formulation is solved opti-mally by using the Gurobi optimization solver [15].
An example set covering matrix and the formulation for the Equity Constrained Proba-bilistic Set Covering Problem EC-PSCP is explained below.
In this example scenario, five fires must be covered by a set of four engine com-panies. The set covering matrix for this scenario is presented in Table 1.
251
-
Table 1: The set covering matrix for example scenario, where columns denote EC =182Engine Company, and rows are fire incidents183
(EC 1) (EC 2) (EC 3) (EC 4)fire 1 1 1 1 0fire 2 1 1 1 0fire 3 1 1 1 1fire 4 0 0 1 0fire 5 0 0 1 1
184
185
The set covering matrixA has entry aij = 1 if an engine company at location (column) j can186reach a fire at location (row) i with required coverage probability p. The constraint set is187illustrated below for this example.188
The Equity Constrained Probabilistic Set Covering Problem (EC-PSCP):189
1. The mathematical program has a variable yjfor each node (potential EC facility).190
2. yj is 1 if an engine company is located at node j, and 0 otherwise (a switch variable).191
3. The math program has a binary variable Xij which is set equal to 1 if fire i is covered192by an engine company located at j.193
4. The math program has a switch variable Si for each fire i. If Si = 1 in the solution,194fire i is NOT covered within the stipulated response time. The variable Si provides195a “pass” that allows the math program to pass on covering fire i (In this model,only196β % of fires are covered within the stipulated response time).197
5. A scenario is defined by a set of locations for fire incidents, drawn from a spatial198distribution.199
6. The mathematical program has five sets of constraints (explained below) in the sce-nario ω.
7. Every new scenario ω will lead to a new mathematical program because the locationsof the fires will change.
200
252
-
OBJECTIVE: Minimize∑|J |j=1 yj201
Illustration for Example: OBJECTIVE: Minimize∑4j=1 yj = y1 + y2 + y3 + y4202
SUBJECT TO FIVE SETS OF CONSTRAINTS:203
Constraint set I (cannot set X ij to 1 unless yj is also 1):204 ∑i
aijX ij ≤ |I |yj for all engine companies j ∈ J
Illustration of Constraint Set I for Example:205
|I| = 5, since there are 5 fires in the example matrix.206
X11 + X21 + X31 ≤ 5y1207
X12 + X22 + X32 ≤ 5y2208
X13 + X23 + X33 + X43 + X53 ≤ 5y3209
X34 + X54 ≤ 5y4210
Constraint set II (either a fire i must be covered by an engine company yj , OR the211“pass” variable Si must be set equal to 1):212 ∑
j
aijyj ≥ [1− Si] for all fires i ∈ I
213
y1 + y2 + y3 ≥ [1− S1] (i.e., cover f ire 1 if S1 = 0)214
y1 + y2 + y3 ≥ [1− S2] (i.e., cover f ire 2 if S2 = 0)215
y1 + y2 + y3 + y4 ≥ [1− S3] (i.e., cover f ire 3 if S3 = 0)216
y3 ≥ [1− S4] (i.e., cover f ire 4 if S4 = 0)217
y3 + y4 ≥ [1− S5] (i.e., cover f ire 5 if S5 = 0)
218
Constraint set III: Enforce the Equity Spread Constraint219 ∑i
aijX ij ≤Max for all engine companies j
220 ∑i
aijX ij ≥Min for all engine companies j
221
222
223
Max − Min ≤ EquitySpread [this is a parameter provided by user]
Illustration of Equity Spread constraints f or example :
X11 + X21 + X31 ≤ Max
253
-
224
X12 + X22 + X32 ≤Max225
X13 + X23 + X33 + X43 + X53 ≤Max226
X34 + X54 ≤Max227
X11 + X21 + X31 ≥Min228
X12 + X22 + X32 ≥Min229
X13 + X23 + X33 + X43 + X53 ≥Min230
X34 + X54 ≥Min231
Max −Min ≤ EquitySpread
Constraint set IV: At least β % of fires must be covered.232
|I |∑i=1
Si ≤ [1−β]×| I |
{equivalently,∑|I |i=1
∑|J |j=1X ij ≥ β ×| I |}233
Illustration of constraint set IV for example, where β = 0.8 is assumed.234
S1 + S2 + S3 + S4 + S5 ≤ { (1− 0.8)× 5} = 1 (This ensures 4 out of 5 fires covered)235
Constraint Set V:236
Declare all variables to be BINARY (0 or 1)237
X ij , yj , Si ∈ {0, 1}
In the example:238
y1, y2, y3, y4 are all binary (0 or 1) variables239
S1..S5 are ALSO binary (0 or 1) variables240
X ij is binary for i = 1, . . . ,5 and j = 1, ..4
241
242
254
-
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
2.4 Combining Multiple Solutions with an Ensemble Algorithm
The Ensemble Algorithm combines the location solutions for each generated sce-nario into a single location solution. Each candidate facility j (in any of the scenario solu-tions) is given a cumulative weight that depends upon the distance of facility j to either a) all fire scenarios or b) to all other selected engine company l ocations. Finally, all the en-gine company locations in any scenario solution are rank-ordered based upon this weight. Engine companies are successively selected from this rank-ordered list in a greedy fashion (i.e., the next best facility is the one that covers the most number of uncovered fires) until the robustness criterion of β % of fires covered is achieved.
2.5 Solving a Constraint Satisfaction Problem using a Genetic Algorithm
When the Ensemble Algorithm is completed, the number of engine companies selected is known along with their precise arcGIS locations. The purpose of Phase III (in Figure 1) is to place resources at these engine companies. Two kinds of resources are considered, pumper engines (that cost about $300,000) and ladder engines (that are more expensive, upwards of $900,000). The goal of the Genetic Algorithm is to determine the type and number of engines at each location. A Genetic Algorithm approach was selected as the search procedure as it offered a natural way to encode the solution space. At least one engine must be placed at each location. In addition, two kinds of constraints must be satisfied by a solution to the Constraint Satisfaction Problem: a) a constraint on the overall cost of engines for the entire city (a budget constraint) and b) every fire must be responded to within 8 minutes by some type of engine (pumper or ladder) and every fire must be responded to with a ladder truck within 18 minutes (these parameters are used for illus-trative purposes only and can be changed by the user of the DSS). The fitness function for the GA is a two-component vector; one component is total cost, and the other is percentage of fires not covered as per constraint b above. A simple algorithm combining crossover and mutation of genes (where a gene is a resource profile for all the selected EC locations) has been designed for this phase. Figures 4 and 5 below provide schematic representations of the Genetic Algorithm steps.
255
-
272
Figure 4: Schematic of Genetic Algorithm.273
Figure 4 provides a schematic representation of the Genetic Algorithm step. The274Gurobi Optimizer [15] is used to solve the probabilistic set covering problem in Phase I.275The optimizer determines the location of the Engine Companies. The Genetic Algorithm276determines how many pumper and ladder trucks to place in each location.277
256
-
278
Figure 5: Schematic for Cross Over and Mutation Operations using Genetic Algorithms279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
Figure 5 provides a schematic representation for cross over and mutation oper-ations in the Genetic Algorithm. There were 100 cross over operations followed by 100 mutation operations for the computational results presented in this paper. However, the user has the ability to set parameters for the number of crossover and mutation opera-tions performed. While in principle, there could be genes that code for a larger number of resources at each Engine Company, in practice, budget limitations would make gene configurations other than the ones discussed in this paper unlikely.
3. Discussion of Computational Results
The methodology presented in the previous section was coded in the Python pro-gramming language. The Equity Constrained Probabilistic Set Covering Problem (EC-PSCP) was solved using the Gurobi optimization package [15]. All computations were per-formed on a 64-bit Lenovo laptop. In conformance with standard practice in the machine learning literature, five training scenarios were created to find engine company solutions and these solutions were validated using five test scenarios (results reported in this section
257
-
are generally averages for the five test scenarios). Please note, the user can also select the294number of training and test scenarios. The discussion of computational results is further295organized into the following sub-sections:296
1. Impact of the robustness parameter β297
2. Impact of engine speed298
3. Solution quality in terms of engine company equity299
4. Impact of coverage probability function assumptions300
5. Illustrative performance of the genetic algorithm.301
3.1 Impact of the Robustness Parameter β302
The parameter β, which is the proportion of fires covered effectively in a solution303is the fundamental measure of “robustness” of any engine company solution. The higher304the β, the more engine companies will be required. The following charts demonstrate the305impact of β on the number of engine companies required and the maximum response time.306
307
Chart 1: Impact of β on the Number of Engine Companies Required308
258
-
309
Chart 2: Impact of β on Maximum Response Times310
259
-
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
Chart 3: Impact of β on Average Response Times
As β increases from 0.9 to 0.99, the average response time decreased from 3.41 to 2.6 minutes. However Charts 1, 2, and 3 together document the price of robustness (# engine companies required more than doubles as β increases from 0.9 to 0.99) and the ben-efit of robustness ( the maximum response time for any fire almost drops by 10 minutes). Policy makers must thoughtfully choose an operating point from these results.
3.2 Impact of Fire Engine Travel SpeedThe speed of travel for any incident may depend upon extraneous factors such as
time of day and traffic conditions. However, most cities have an inherently latent capacity for allowing fire trucks to move about the city with a certain speed. Chart 4, below, shows a steep increase in the number of engine companies is required for travel speeds less than 30 mph. In this research, rather than view the fire engine travel speed as a given of the environment, it is assumed that policy makers have a limited ability to influence the speed of travel for emergency response (e.g., by levying extreme fines for obstructors).
260
-
326
327
328
329
330
331
332
333
334
335
336
337
338
Chart 4: Impact of Engine Speed on Number of Engine Companies Required
3.3 Equity for Engine Company WorkloadsA good solution for the engine company location problem must pay attention to
the workloads allocated to various fire engine companies (some companies cannot remain idle too much of the time). Equity considerations can be explicitly added to the EC-PSCP, but this also makes the problem more difficult to solve. Future research should also ex-plore improving the equity with Phase III (local search for better solutions). Chart 5 below indicates a discrepancy of about 10% (between the busiest and most idle engine compa-nies) when EC-PSCP is solved without any constraints on equitable workloads. Moreover, equity is harder to achieve for higher values of β.
261
-
339
340
Chart 5: Impact of Equity Spread on Engine Company Workloads as β Increases341
342
However, EC-PSCP can also be solved parametrically by varying the “Equity343Spread” limit in the formulation. Clearly, the number of engine companies needed, as well344as the total system cost, can be expected to increase as the Equity Spread limit decreases.345The behavior of the system with decreasing Equity Spread is characterized in Charts 6 and3467 below.347
262
-
348
Chart 6: Impact of Equity Spread on Number of Engine Companies Required349
263
-
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
Chart 7: Impact of Equity Spread on Total Cost
For low levels of Equity Spreads, the system cost increases by as much as $1M for each percentage point reduction in the Equity Spread. From these plots, it can be observed that the increase in system cost depends upon the current level of the equity spread. So a decrease in Equity Spread from 15 to 14 has a much smaller impact on system cost as compared to a decrease from 2 to 1.
3.4 Impact of Coverage Probability Function Assumptions
The user-specified coverage probability function of Chart 1 is a key determinant of the solution quality. In particular, tmin and tmax in Chart 1 influence the form of the cov-erage probability function and in turn the number of engine companies needed. In some cases, the coverage probability function may be the subjective opinion of an expert user of the DSS. For this reason, the impact of changing the form of the coverage probability func-tion is studied below in Chart 8. The National Fire Protection Association (NFPA) standard
264
-
calls for the first due fire engines to arrive on scene within 5 minutes and 20 seconds after365being dispatched for 90 percent of their runs. As expected, response times lower than366about 6 minutes impose an enormous cost on the system. [18]367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
Chart 8: Impact of Response Time Standards on Number of Engine Companies
3.5 Illustrative Performance of the Genetic Algorithm (GA)
Phase III of the solution procedure employs a GA to place resources (pumper and ladder trucks) at each chosen engine company location. Only resource allocations are changed in this phase and engine company locations themselves are fixed. The GA uses a two-dimensional fitness function. The first dimension is the budget used. Given that there must be at least one engine at each location, a lower bound for the cost is the number of engine companies multiplied by the cost of a pumper (less expensive) truck. All resource configuration c osts a re e xpressed a s a n i ndex w ith r espect t o t his b ase c ost ( i.e., a cost of 150 means the resource configuration is 50% more expensive than the cost of placing one pumper truck at all stations). Likewise, the second component of the fitness function represents the percentage of fires that do not satisfy the constraint “first response (by any type of truck) within 8 minutes and a ladder truck available within 18 minutes”. The GA searched for a resource configuration with a cost index less than 200 and the percentage of infeasible fires less than 10% (For illustration only: the user can tune feasibility parameters for genes). Chart 9 below illustrates how these two GA fitness attributes trade off in the set of feasible solutions found. The computational experiments indicate that solutions feasible to these two constraints are hard to find (less than 10% of the population members generated by GA were “feasible”).
265
-
389
390
Chart 9: GA Performance: System Cost vs. Percentage of Fires Not Covered391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
Finally, a check was made for overfitting. I f t he β c alculated f or t he t est data sets is significantly lower than the β stipulated for the training data s ets, overfitting has occurred. It was validated that the model was not overfitting the d ata. There were some “negative” results in the computations however. For instance, in the Ensemble Algorithm, it did not matter whether facilities were combined based upon distance to fires in scenarios, or distance to other facilities chosen in other solutions. Moreover, the number of candidate engine company locations that the DSS started with did not matter beyond a point. Most of the reported computations started with 625 candidate engine company locations.
4. Conclusions, Public Policy Implications, and Future Research Opportunities
4.1 ConclusionsThis paper develops a robust optimization approach for locating fire engine com-
panies. The main dimension of robustness addressed is the spatial uncertainty of fire in-cident locations. For this problem, the current paper provides a solution algorithm to find the minimum number of engine companies needed so that β % of fires can be covered with
266
-
probability p. The main contribution is the development of a Probabilistic Set Covering407Model formulation for this problem. In addition to the EC-PSCP, the paper also develops408a method to combine solutions from different scenarios (the Ensemble algorithm) and a409local search procedure for placing resources at chosen fire engine company locations.410
4.2 Public Policy Implications411
The solution algorithm and associated computational results raise several impor-412tant public policy issues:413
1. One of the responsibilities of city government is to choose an appropriate level of414coverage β. In fact, the required number of engine companies more than doubles415as β increases from 90 to 99%. Policy makers should carefully evaluate the trade-off416between the increased cost of opening additional locations and the real benefits from417covering the last few percentiles.418
2. As engine speed increases, the number of engine companies required decreases (as419expected). The data also seem to indicate that EMS responders must try to achieve420an engine speed of at least 30 mph. For engine speeds < 30 mph, there is again a421steep increase in the number of engine companies required. Given that this pa-422rameter has the greatest impact on the number of engine companies needed, policy423makers should consider policy options that can increase fire engine travel speeds.424Options such as special lanes for fire engine travel (like what is done for mass transit425buses already) and higher fines on the roads for obstructing fire engine travel must426be considered.427
3. Special attention must be given to engine company workload equity. The computa-428tional results indicate that there might be as much as a 10% difference in the propor-429tion of fires tackled by the busiest engine company and the engine company with the430smallest workload. Territory design for engine companies (to attain equitable work-431loads) is an important extension of this project, as disparity may cause difficulties432with labor unions or contract workers.433
4. Finally, achieving response times of less than 5 minutes is extraordinarily difficult.434There is a steep increase in the number of engine companies required for achiev-435ing average response times less than 5 minutes. Policy makers should consider436options wherein a sufficient amount of education and equipment is provided locally437(e.g., fire extinguishers) at building sites, so that local residents can contain the im-438pact of the fire for about 6-8 minutes. If this 6-8-minute time threshold can be man-439aged locally, the city can also drastically lower the costs of opening more fire engine440companies.441
442
443
444
445
446
447
4.3 Future Research Opportunities
The current research also has some modeling limitations. Travel times are com-puted ignoring conditions like traffic or time of day. Some fires require multiple response units and it may not be sufficient to dispatch just the closest unit. Some fires m ay also require specific equipment such as ladder companies that may not be available at the clos-est facility (this issue is partly addressed by the Genetic Algorithm). Fire engines may
267
-
also be unavailable due to external circumstances such as maintenance and maintenance448plans must be factored into developing engine company locations. In terms of model-449ing enhancements, this exercise can be repeated with GIS mapping tools to formulate the450same model at a more fine-grained level (e.g., include details of one-way streets, traffic451lights/intersections). Finally, the model developed herein has a rich set of other appli-452cations such as ambulance shelter location, police patrol improvement and logistics for453emergency response (e.g., for hurricanes such as Katrina), where spatial demand uncer-454tainty considerations are required.455
5. References456[1] R. ALANIS, A. INGOLFSSON, and B. KOLFAL, A Markov Chain Model for an EMS457
System with Repositioning, Production and Operations Management, Vol. 22, No. 1, (2013),458pp. 216–231.459
460
[2] O. BARON, J. MILNER, and H. NASERALDIN, Facility location: a robust optimization461approach, Production and Operations Management, Vol. 20, Issue 5, (2011), pp. 772-785.462
463
[3] F.P. BOSCOE, K.A. HENRY, and M.S. ZDEB, A Nationwide Comparison of Driving Dis-464tance Versus Straight-Line Distance to Hospitals, Prof Geogr. , 64(2), Apr 1 (2012).465
466
[4] F.P. BOSCOE, K.A. HENRY, and M.S. ZDEB, A Nationwide Comparison of Driving Dis-467tance Versus Straight-Line Distance to Hospitals, Prof Geogr. , 64(2), Apr 1 (2012).468
469
[5] S. BUDGE, A. INGOLFSSON, and D. ZEROM, Empirical Analysis of Ambulance Travel470Times: The Case of Calgary Emergency Medical Services, Management Science, Vol. 56,471Issue 4, (2010) pp. 716 – 723.472
473
[6] T.C.Y. CHAN, D. DEMIRTAS, and R. KWON, Optimizing the Deployment of Public Ac-474cess Defibrillators, published as Articles in Advance, Management Science, January (2016).475
476
[7] T.CUI, Y.OUYANG, and Z.M. SHEN, Reliable facility location design under the risk of477disruptions, Operations Research, Vol. 58, Issue 4, (2010).478
479
[8] L.V. GREEN and P.J. KOLESAR, Improving emergency response with management sci-480ence, Management Science, Vol. 50, No. 8, (2004), pp. 1001 – 1014.481
482
268
-
[9] S.O. KIMBROUGH, A. KUO, L.H. CHUIN, F.H. MURPHY, and D.H. WOOD, Solution483pluralism and metaheuristics, The IX Metaheuristics International Conference, Udine, Italy,484July 25-28, (2011).485
486
[10] P. KOLESAR and E.H. BLUM, Square root laws for fire engine response distances,487Management Science, 19(12), (1973), pp. 1368-1378.488
489
[11] P. KOLESAR and W.E. WALKER, An algorithm for the dynamic relocation of fire com-490panies, Operations Research, 22(2), (1974), pp. 249-274.491
492
[12] R.C. LARSON and T.F. RICH., Travel time analysis of New York city police cars, Inter-493faces, 17(2), (1987), pp. 15-20.494
495
[13] D.W. SCOTT, Multivariate Density Estimation: Theory, Practice and Visualization, Wiley496Series in Probability and Statistics, Vol. 383 (John Wiley Sons, Hoboken, NJ) (2009).497
498
[14] Four children killed in 3 alarm SW Philly fire, www.phillyfirenews.com/2014/07/49905/four-4-kids-killed-3-alarm-sw-philly-fire/, downloaded on September 16, (2016).500
501
[15] Gurobi Optimizer, http://www.gurobi.com/, downloaded on September 16, (2016).502503
[16] Map of Philadelphia Neighborhoods, http://www.arcgis.com/home/webmap/viewer.504html?webmap=d21eb07ec54a4dcb9f80d057fe7c3cd6, downloaded on September 16, (2016).505
506
[17] Three alarm fire at apartment building in west Philadelphia, Philadelphia.cbslocal.507com/2016/06/12/three-alarm-fire-at-apartment-building-in-west-philadelphia,508downloaded on September 16, (2016).509
510
[18] PHILLY REPORT CRITICAL OF RESPONSE TIMES, http://www.philadelphiacontroller.511org/publications/FireDepartment_ResponseTime_February2016.pdf, downloaded on512July 18, (2016).513
514
515
269
www.phillyfirenews.com/2014/07/05/four-4-kids-killed-3-alarm-sw-philly-fire/www.phillyfirenews.com/2014/07/05/four-4-kids-killed-3-alarm-sw-philly-fire/www.phillyfirenews.com/2014/07/05/four-4-kids-killed-3-alarm-sw-philly-fire/http://www.gurobi.com/http://www.arcgis.com/home/webmap/viewer.html?webmap=d21eb07ec54a4dcb9f80d057fe7c3cd6http://www.arcgis.com/home/webmap/viewer.html?webmap=d21eb07ec54a4dcb9f80d057fe7c3cd6http://www.arcgis.com/home/webmap/viewer.html?webmap=d21eb07ec54a4dcb9f80d057fe7c3cd6Philadelphia.cbslocal.com/2016/06/12/three-alarm-fire-at-apartment-building-in-west-philadelphiaPhiladelphia.cbslocal.com/2016/06/12/three-alarm-fire-at-apartment-building-in-west-philadelphiaPhiladelphia.cbslocal.com/2016/06/12/three-alarm-fire-at-apartment-building-in-west-philadelphiahttp://www.philadelphiacontroller.org/publications/FireDepartment_ResponseTime_February2016.pdfhttp://www.philadelphiacontroller.org/publications/FireDepartment_ResponseTime_February2016.pdfhttp://www.philadelphiacontroller.org/publications/FireDepartment_ResponseTime_February2016.pdf